자산 간 상관관계 기반 지도학습 특성선택을 활용한 거시 자산 ETF 가격 예측 모형 연구

초록

The interdependence structure among financial assets may contain useful information for price prediction; however, correlation features derived from numerous asset pairs entail overfitting risks due to high dimensionality. This study analyzes the effect of cross-asset correlation features on next-day closing price prediction using daily data of 11 macro ETFs from 2011 to 2025. Applying F-statistic-based univariate feature selection to 55 non-redundant features derived from Pearson correlation and variation of information, we find that a configuration selecting only five key features yields a statistically significant improvement over the baseline model. The predictive contribution of correlation features is relatively higher during high-volatility regimes including the COVID-19 pandemic period. However, a simple persistence forecast attains lower error than all configurations, so the contribution of correlation features should be interpreted as marginal information added to the baseline rather than as absolute predictive accuracy. This study suggests that simple supervised feature selection can be of limited but measurable use in utilizing high-dimensional correlation features.

키워드